This document attempts to help you work with the Blender API in areas that can be troublesome and avoid practices that are known to cause instability.

Using Operators#

Blender’s operators are tools for users to access, that can be accessed with Python too which is very useful. Still operators have limitations that can make them cumbersome to script.

The main limits are:

  • Can’t pass data such as objects, meshes or materials to operate on (operators use the context instead).

  • The return value from calling an operator is the success (if it finished or was canceled), in some cases it would be more logical from an API perspective to return the result of the operation.

  • Operators’ poll function can fail where an API function would raise an exception giving details on exactly why.

Why does an operator’s poll fail?#

When calling an operator it gives an error like this:

>>> bpy.ops.action.clean(threshold=0.001)
RuntimeError: Operator bpy.ops.action.clean.poll() failed, context is incorrect

Which raises the question as to what the correct context might be?

Typically operators check for the active area type, a selection or active object they can operate on, but some operators are more strict when they run. In most cases you can figure out what context an operator needs by examining how it’s used in Blender and thinking about what it does.

If you’re still stuck, unfortunately, the only way to eventually know what is causing the error is to read the source code for the poll function and see what it is checking. For Python operators it’s not so hard to find the source since it’s included with Blender and the source file and line is included in the operator reference docs. Downloading and searching the C code isn’t so simple, especially if you’re not familiar with the C language but by searching the operator name or description you should be able to find the poll function with no knowledge of C.


Blender does have the functionality for poll functions to describe why they fail, but it’s currently not used much, if you’re interested to help improve the API feel free to add calls to bpy.types.Operator.poll_message_set (CTX_wm_operator_poll_msg_set in C) where it’s not obvious why poll fails, e.g:

>>> bpy.ops.gpencil.draw()
RuntimeError: Operator bpy.ops.gpencil.draw.poll() Failed to find Grease Pencil data to draw into

The operator still doesn’t work!#

Certain operators in Blender are only intended for use in a specific context, some operators for example are only called from the properties editor where they check the current material, modifier or constraint.

Examples of this are:

Another possibility is that you are the first person to attempt to use this operator in a script and some modifications need to be made to the operator to run in a different context. If the operator should logically be able to run but fails when accessed from a script it should be reported to the bug tracker.

Stale Data#

No updates after setting values#

Sometimes you want to modify values from Python and immediately access the updated values, e.g: Once changing the objects bpy.types.Object.location you may want to access its transformation right after from bpy.types.Object.matrix_world, but this doesn’t work as you might expect. There are similar issues with changes to the UI, that are covered in the next section.

Consider the calculations that might contribute to the object’s final transformation, this includes:

  • Animation function curves.

  • Drivers and their Python expressions.

  • Constraints

  • Parent objects and all of their F-Curves, constraints, etc.

To avoid expensive recalculations every time a property is modified, Blender defers the evaluation until the results are needed. However, while the script runs you may want to access the updated values. In this case you need to call bpy.types.ViewLayer.update after modifying values, for example:

bpy.context.object.location = 1, 2, 3

Now all dependent data (child objects, modifiers, drivers, etc.) have been recalculated and are available to the script within the active view layer.

No updates after changing UI context#

Similar to the previous issue, some changes to the UI may also not have an immediate effect. For example, setting bpy.types.Window.workspace doesn’t seem to cause an observable effect in the immediately following code (bpy.types.Window.workspace is still the same), but the UI will in fact reflect the change. Some of the properties that behave that way are:

Such changes impact the UI, and with that the context (bpy.context) quite drastically. This can break Blender’s context management. So Blender delays this change until after operators have run and just before the UI is redrawn, making sure that context can be changed safely.

If you rely on executing code with an updated context this can be worked around by executing the code in a delayed fashion as well. Possible options include:

It’s also possible to depend on drawing callbacks although these should generally be avoided as failure to draw a hidden panel, region, cursor, etc. could cause your script to be unreliable

Can I redraw during script execution?#

The official answer to this is no, or… “You don’t want to do that”. To give some background on the topic:

While a script executes, Blender waits for it to finish and is effectively locked until it’s done; while in this state Blender won’t redraw or respond to user input. Normally this is not such a problem because scripts distributed with Blender tend not to run for an extended period of time, nevertheless scripts can take a long time to complete and it would be nice to see progress in the viewport.

Tools that lock Blender in a loop redraw are highly discouraged since they conflict with Blender’s ability to run multiple operators at once and update different parts of the interface as the tool runs.

So the solution here is to write a modal operator, which is an operator that defines a modal() function, See the modal operator template in the text editor. Modal operators execute on user input or setup their own timers to run frequently, they can handle the events or pass through to be handled by the keymap or other modal operators. Examples of modal operators are Transform, Painting, Fly Navigation and File Select.

Writing modal operators takes more effort than a simple for loop that contains draw calls but is more flexible and integrates better with Blender’s design.

Ok, Ok! I still want to draw from Python

If you insist – yes it’s possible, but scripts that use this hack will not be considered for inclusion in Blender and any issue with using it will not be considered a bug, there is also no guaranteed compatibility in future releases.

bpy.ops.wm.redraw_timer(type='DRAW_WIN_SWAP', iterations=1)

Modes and Mesh Access#

When working with mesh data you may run into the problem where a script fails to run as expected in Edit-Mode. This is caused by Edit-Mode having its own data which is only written back to the mesh when exiting Edit-Mode.

A common example is that exporters may access a mesh through (a bpy.types.Mesh) when the user is in Edit-Mode, where the mesh data is available but out of sync with the edit mesh.

In this situation you can…

  • Exit Edit-Mode before running the tool.

  • Explicitly update the mesh by calling bmesh.types.BMesh.to_mesh.

  • Modify the script to support working on the edit-mode data directly, see: bmesh.from_edit_mesh.

  • Report the context as incorrect and only allow the script to run outside Edit-Mode.

N-Gons and Tessellation#

Since 2.63 n-gons are supported, this adds some complexity since in some cases you need to access triangles still (some exporters for example).

There are now three ways to access faces:

For the purpose of the following documentation, these will be referred to as polygons, loop triangles and BMesh-faces respectively.

Faces with five or more sides will be referred to as ngons.

Support Overview#






Poor (inflexible)

Unusable (read-only).



Poor (inflexible)

Unusable (read-only).



Good (n-gon support)

Good (When n-gons cannot be used)

Good (n-gons, extra memory overhead)


Using the bmesh API is completely separate API from bpy, typically you would use one or the other based on the level of editing needed, not simply for a different way to access faces.


All three data types can be used for face creation:

  • Polygons are the most efficient way to create faces but the data structure is very rigid and inflexible, you must have all your vertices and faces ready and create them all at once. This is further complicated by the fact that each polygon does not store its own vertices, rather they reference an index and size in bpy.types.Mesh.loops which are a fixed array too.

  • BMesh-faces are most likely the easiest way to create faces in new scripts, since faces can be added one by one and the API has features intended for mesh manipulation. While bmesh.types.BMesh uses more memory it can be managed by only operating on one mesh at a time.


Editing is where the three data types vary most.

  • Polygons are very limited for editing, changing materials and options like smooth works, but for anything else they are too inflexible and are only intended for storage.

  • Loop-triangles should not be used for editing geometry because doing so will cause existing n-gons to be tessellated.

  • BMesh-faces are by far the best way to manipulate geometry.


All three data types can be used for exporting, the choice mostly depends on whether the target format supports n-gons or not.

  • Polygons are the most direct and efficient way to export providing they convert into the output format easily enough.

  • Loop-triangles work well for exporting to formats which don’t support n-gons, in fact this is the only place where their use is encouraged.

  • BMesh-Faces can work for exporting too but may not be necessary if polygons can be used since using BMesh gives some overhead because it’s not the native storage format in Object-Mode.

Edit Bones, Pose Bones, Bone… Bones#

Armature Bones in Blender have three distinct data structures that contain them. If you are accessing the bones through one of them, you may not have access to the properties you really need.


In the following examples bpy.context.object is assumed to be an armature object.

Edit Bones# contains an edit bones; to access them you must set the armature mode to Edit-Mode first (edit bones do not exist in Object or Pose-Mode). Use these to create new bones, set their head/tail or roll, change their parenting relationships to other bones, etc.

Example using bpy.types.EditBone in armature Edit-Mode which is only possible in Edit-Mode:

>>>["Bone"].head = Vector((1.0, 2.0, 3.0))

This will be empty outside of Edit-Mode:

>>> mybones = bpy.context.selected_editable_bones

Returns an edit bone only in Edit-Mode:

>>> bpy.context.active_bone

Bones (Object-Mode)# contains bones. These live in Object-Mode, and have various properties you can change, note that the head and tail properties are read-only.

Example using bpy.types.Bone in Object or Pose-Mode returning a bone (not an edit bone) outside of Edit-Mode:

>>> bpy.context.active_bone

This works, as with Blender the setting can be edited in any mode:

>>>["Bone"].use_deform = True

Accessible but read-only:

>>> tail =["Bone"].tail

Pose Bones#

bpy.context.object.pose.bones contains pose bones. This is where animation data resides, i.e. animatable transformations are applied to pose bones, as are constraints and IK-settings.

Examples using bpy.types.PoseBone in Object or Pose-Mode:

# Gets the name of the first constraint (if it exists)

# Gets the last selected pose bone (Pose-Mode only)


Notice the pose is accessed from the object rather than the object data, this is why Blender can have two or more objects sharing the same armature in different poses.


Strictly speaking pose bones are not bones, they are just the state of the armature, stored in the bpy.types.Object rather than the bpy.types.Armature, yet the real bones are accessible from the pose bones via bpy.types.PoseBone.bone.

Armature Mode Switching#

While writing scripts that deal with armatures you may find you have to switch between modes, when doing so take care when switching out of Edit-Mode not to keep references to the edit bones or their head/tail vectors. Further access to these will crash Blender so it’s important that the script clearly separates sections of the code which operate in different modes.

This is mainly an issue with Edit-Mode since pose data can be manipulated without having to be in Pose-Mode, yet for operator access you may still need to enter Pose-Mode.

Data Names#

Naming Limitations#

A common mistake is to assume newly created data is given the requested name. This can cause bugs when you add data (normally imported) then reference it later by name:

# normally some code, function calls...[meshid]

Or with name assignment: = objname

# normally some code, function calls...
obj =[objname]

Data names may not match the assigned values if they exceed the maximum length, are already used or an empty string.

It’s better practice not to reference objects by names at all, once created you can store the data in a list, dictionary, on a class, etc; there is rarely a reason to have to keep searching for the same data by name.

If you do need to use name references, it’s best to use a dictionary to maintain a mapping between the names of the imported assets and the newly created data, this way you don’t run this risk of referencing existing data from the blend-file, or worse modifying it.

# typically declared in the main body of the function.
mesh_name_mapping = {}

mesh =
mesh_name_mapping[meshid] = mesh

# normally some code, or function calls...

# use own dictionary rather than
mesh = mesh_name_mapping[meshid]

Library Collisions#

Blender keeps data names unique ( so you can’t name two objects, meshes, scenes, etc., the same by accident. However, when linking in library data from another blend-file naming collisions can occur, so it’s best to avoid referencing data by name at all.

This can be tricky at times and not even Blender handles this correctly in some cases (when selecting the modifier object for e.g. you can’t select between multiple objects with the same name), but it’s still good to try avoiding these problems in this area. If you need to select between local and library data, there is a feature in members to allow for this.

# typical name lookup, could be local or library.
obj =["my_obj"]

# library object name look up using a pair
# where the second argument is the library path matching bpy.types.Library.filepath
obj =["my_obj", "//my_lib.blend"]

# local object name look up using a pair
# where the second argument excludes library data from being returned.
obj =["my_obj", None]

# both the examples above also works for 'get'
obj ="my_obj", None))

Relative File Paths#

Blender’s relative file paths are not compatible with standard Python modules such as sys and os. Built-in Python functions don’t understand Blender’s // prefix which denotes the blend-file path.

A common case where you would run into this problem is when exporting a material with associated image paths:

>>> bpy.path.abspath(image.filepath)

When using Blender data from linked libraries there is an unfortunate complication since the path will be relative to the library rather than the open blend-file. When the data block may be from an external blend-file pass the library argument from the bpy.types.ID.

>>> bpy.path.abspath(image.filepath, library=image.library)

These returns the absolute path which can be used with native Python modules.

Unicode Problems#

Python supports many different encodings so there is nothing stopping you from writing a script in latin1 or iso-8859-15. See PEP 263.

However, this complicates matters for Blender’s Python API because .blend files don’t have an explicit encoding. To avoid the problem for Python integration and script authors we have decided that all strings in blend-files must be UTF-8, ASCII compatible. This means assigning strings with different encodings to an object name, for instance, will raise an error.

Paths are an exception to this rule since the existence of non-UTF-8 paths on the user’s file system cannot be ignored. This means seemingly harmless expressions can raise errors, e.g:

>>> print(
UnicodeEncodeError: 'ascii' codec can't encode characters in position 10-21: ordinal not in range(128)
>>> =
Traceback (most recent call last):
  File "<blender_console>", line 1, in <module>
TypeError: bpy_struct: item.attr= val: expected a string type, not str

Here are two ways around file-system encoding issues:

>>> print(repr(
>>> import os
>>> filepath_bytes = os.fsencode(
>>> filepath_utf8 = filepath_bytes.decode('utf-8', "replace")
>>> = filepath_utf8

Unicode encoding/decoding is a big topic with comprehensive Python documentation, to keep it short about encoding problems – here are some suggestions:

  • Always use UTF-8 encoding or convert to UTF-8 where the input is unknown.

  • Avoid manipulating file paths as strings directly, use os.path functions instead.

  • Use os.fsencode() or os.fsdecode() instead of built-in string decoding functions when operating on paths.

  • To print paths or to include them in the user interface use repr(path) first or "%r" % path with string formatting.


Sometimes it’s preferable to avoid string encoding issues by using bytes instead of Python strings, when reading some input it’s less trouble to read it as binary data though you will still need to decide how to treat any strings you want to use with Blender, some importers do this.

Strange Errors when Using the ‘Threading’ Module#

Python threading with Blender only works properly when the threads finish up before the script does, for example by using threading.join().

Here is an example of threading supported by Blender:

import threading
import time

def prod():
    print(threading.current_thread().name, "Starting")

    # do something vaguely useful
    import bpy
    from mathutils import Vector
    from random import random

    prod_vec = Vector((random() - 0.5, random() - 0.5, random() - 0.5))
    print("Prodding", prod_vec)["Cube"].location += prod_vec
    time.sleep(random() + 1.0)
    # finish

    print(threading.current_thread().name, "Exiting")

threads = [threading.Thread(name="Prod %d" % i, target=prod) for i in range(10)]

print("Starting threads...")

for t in threads:

print("Waiting for threads to finish...")

for t in threads:

This an example of a timer which runs many times a second and moves the default cube continuously while Blender runs (Unsupported).

def func():
    import bpy['Cube'].location.x += 0.05

def my_timer():
    from threading import Timer
    t = Timer(0.1, my_timer)


Use cases like the one above which leave the thread running once the script finishes may seem to work for a while but end up causing random crashes or errors in Blender’s own drawing code.

So far, no work has been done to make Blender’s Python integration thread safe, so until it’s properly supported, it’s best not make use of this.


Python threads only allow concurrency and won’t speed up your scripts on multiprocessor systems, the subprocess and multiprocess modules can be used with Blender to make use of multiple CPUs too.

Help! My script crashes Blender#

TL;DR Do not keep direct references to Blender data (of any kind) when modifying the container of that data, and/or when some undo/redo may happen (e.g. during modal operators execution…). Instead, use indices (or other data always stored by value in Python, like string keys…), that allow you to get access to the desired data.

Ideally it would be impossible to crash Blender from Python, however, there are some problems with the API where it can be made to crash. Strictly speaking this is a bug in the API but fixing it would mean adding memory verification on every access since most crashes are caused by the Python objects referencing Blender’s memory directly, whenever the memory is freed or re-allocated, further Python access to it can crash the script. But fixing this would make the scripts run very slow, or writing a very different kind of API which doesn’t reference the memory directly.

Here are some general hints to avoid running into these problems:

  • Be aware of memory limits, especially when working with large lists since Blender can crash simply by running out of memory.

  • Many hard to fix crashes end up being because of referencing freed data, when removing data be sure not to hold any references to it.

  • Re-allocation can lead to the same issues (e.g. if you add a lot of items to some Collection, this can lead to re-allocating the underlying container’s memory, invalidating all previous references to existing items).

  • Modules or classes that remain active while Blender is used, should not hold references to data the user may remove, instead, fetch data from the context each time the script is activated.

  • Crashes may not happen every time, they may happen more on some configurations or operating systems.

  • Be careful with recursive patterns, those are very efficient at hiding the issues described here.

  • See last subsection about Unfortunate Corner Cases for some known breaking exceptions.


To find the line of your script that crashes you can use the faulthandler module. See the Faulthandler docs.

While the crash may be in Blender’s C/C++ code, this can help a lot to track down the area of the script that causes the crash.


Some container modifications are actually safe, because they will never re-allocate existing data (e.g. linked lists containers will never re-allocate existing items when adding or removing others).

But knowing which cases are safe and which aren’t implies a deep understanding of Blender’s internals. That’s why, unless you are willing to dive into the RNA C implementation, it’s simpler to always assume that data references will become invalid when modifying their containers, in any possible way.

Do not:

class TestItems(bpy.types.PropertyGroup):
    name: bpy.props.StringProperty()

bpy.types.Scene.test_items = bpy.props.CollectionProperty(type=TestItems)

first_item = bpy.context.scene.test_items.add()
for i in range(100):

# This is likely to crash, as internal code may re-allocate
# the whole container (the collection) memory at some point. = "foobar"


class TestItems(bpy.types.PropertyGroup):
    name: bpy.props.StringProperty()

bpy.types.Scene.test_items = bpy.props.CollectionProperty(type=TestItems)

first_item = bpy.context.scene.test_items.add()
for i in range(100):

# This is safe, we are getting again desired data *after*
# all modifications to its container are done.
first_item = bpy.context.scene.test_items[0] = "foobar"


For safety, you should assume that undo and redo always invalidates all bpy.types.ID instances (Object, Scene, Mesh, Light, etc.), as well obviously as all of their sub-data.

This example shows how you can tell undo changes the memory locations:

>>> hash(bpy.context.object)
>>> hash(bpy.context.object)

Delete the active object, then undo:

>>> hash(bpy.context.object)

As suggested above, simply not holding references to data when Blender is used interactively by the user is the only way to make sure that the script doesn’t become unstable.


Modern undo/redo system does not systematically invalidate all pointers anymore. Some data (in fact, most data, in typical cases), which were detected as unchanged for a particular history step, may remain unchanged and hence their pointers may remain valid.

Be aware that if you want to take advantage of this behavior for some reason, there is no guarantee of any kind that it will be safe and consistent. Use it at your own risk.

Undo & Library Data#

One of the advantages with Blender’s library linking system that undo can skip checking changes in library data since it is assumed to be static. Tools in Blender are not allowed to modify library data. But Python does not enforce this restriction.

This can be useful in some cases, using a script to adjust material values for example. But it’s also possible to use a script to make library data point to newly created local data, which is not supported since a call to undo will remove the local data but leave the library referencing it and likely crash.

So it’s best to consider modifying library data an advanced usage of the API and only to use it when you know what you’re doing.

Abusing RNA property callbacks#

Python-defined RNA properties can have custom callbacks. Trying to perform complex operations from there, like calling an operator, may work, but is not officially recommended nor supported.

Main reason is that those callback should be very fast, but additionally, it may for example create issues with undo/redo system (most operators store an history step, and editing an RNA property does so as well), trigger infinite update loops, and so on.

Edit-Mode / Memory Access#

Switching mode bpy.ops.object.mode_set(mode='EDIT') or bpy.ops.object.mode_set(mode='OBJECT') will re-allocate objects data, any references to a meshes vertices/polygons/UVs, armatures bones, curves points, etc. cannot be accessed after switching mode.

Only the reference to the data itself can be re-accessed, the following example will crash.

mesh =
polygons = mesh.polygons

# this will crash

So after switching mode you need to re-access any object data variables, the following example shows how to avoid the crash above.

mesh =
polygons = mesh.polygons

# polygons have been re-allocated
polygons = mesh.polygons

These kinds of problems can happen for any functions which re-allocate the object data but are most common when switching mode.

Array Re-Allocation#

When adding new points to a curve or vertices/edges/polygons to a mesh, internally the array which stores this data is re-allocated.

point =[0].bezier_points[0][0].bezier_points.add()

# this will crash! = 1.0, 2.0, 3.0

This can be avoided by re-assigning the point variables after adding the new one or by storing indices to the points rather than the points themselves.

The best way is to sidestep the problem altogether by adding all the points to the curve at once. This means you don’t have to worry about array re-allocation and it’s faster too since re-allocating the entire array for every added point is inefficient.

Removing Data#

Any data that you remove shouldn’t be modified or accessed afterwards, this includes: F-Curves, drivers, render layers, timeline markers, modifiers, constraints along with objects, scenes, collections, bones, etc.

The remove() API calls will invalidate the data they free to prevent common mistakes. The following example shows how this precaution works:

mesh ="MyMesh")
# normally the script would use the mesh here...
print(  # <- give an exception rather than crashing:

# ReferenceError: StructRNA of type Mesh has been removed

But take care because this is limited to scripts accessing the variable which is removed, the next example will still crash:

mesh ="MyMesh")
vertices = mesh.vertices
print(vertices)  # <- this may crash

Unfortunate Corner Cases#

Besides all expected cases listed above, there are a few others that should not be an issue but, due to internal implementation details, currently are:

Collection Objects#

Changing: Object.hide_viewport, Object.hide_select or Object.hide_render will trigger a rebuild of Collection caches, thus breaking any current iteration over Collection.all_objects.

Do not:

# `all_objects` is an iterator. Using it directly while performing operations on its members that will update
# the memory accessed by the `all_objects` iterator will lead to invalid memory accesses and crashes.
for object in["Collection"].all_objects:
     object.hide_viewport = True


# `all_objects[:]` is an independent list generated from the iterator. As long as no objects are deleted,
# its content will remain valid even if the data accessed by the `all_objects` iterator is modified.
for object in["Collection"].all_objects[:]:
     object.hide_viewport = True

Data-Blocks Renaming During Iteration#

Data-blocks accessed from are sorted when their name is set. Any loop that iterates of a data such as for example, and sets the objects name must get all items from the iterator first (typically by converting to a list or tuple) to avoid missing some objects and iterating over others multiple times.


Some Python modules will call sys.exit() themselves when an error occurs, while not common behavior this is something to watch out for because it may seem as if Blender is crashing since sys.exit() will close Blender immediately.

For example, the argparse module will print an error and exit if the arguments are invalid.

An dirty way of troubleshooting this is to set sys.exit = None and see what line of Python code is quitting, you could of course replace sys.exit with your own function but manipulating Python in this way is bad practice.