anomalyco/opencode
1
0
Fork 0
mirror of https://github.com/anomalyco/opencode.git synced 2026-07-22 18:25:32 -04:00
Code Issues 5.9k Packages Projects Releases 19 Wiki Activity

Feature Request: Global default tools configuration for agents #1895

New Issue
Closed
opened 2026-02-16 17:33:06 -05:00 by yindo · 13 comments
yindo commented 2026-02-16 17:33:06 -05:00
Owner
Copy Link
Copy Source

Originally created by @calvinalkan on GitHub (Sep 30, 2025).

Originally assigned to: @rekram1-node on GitHub.

Problem

When adding MCP servers or as new built-in tools are added, they're automatically
available to all agents. This causes:

  • Context window bloat from unused tool prompts
  • Need to manually disable tools in every agent configuration
  • No way to set a "baseline" of tools that agents inherit

Proposed Solution

Add a global defaultTools array that defines which tools are available to all
agents by default:

{
  "$schema": "https://opencode.ai/config.json",
  "defaultTools": ["read", "grep", "glob", "list"],
  "agent": {
    "myagent": {
      "description": "Custom agent",
      "tools": {
        "webfetch": true
      }
    }
  }
}

Behavior

• If defaultTools not set: Current behavior (all tools enabled by default) -
non-breaking
• If defaultTools is set: Only these tools are available by default, agents can
add more via their tools config
• Users can set "defaultTools": [] (empty array): Zero tools by default, opt-in
per agent

This allows clean agent contexts without repetitive configuration across all
agents.

Originally created by @calvinalkan on GitHub (Sep 30, 2025). Originally assigned to: @rekram1-node on GitHub. ## Problem When adding MCP servers or as new built-in tools are added, they're automatically available to all agents. This causes: - Context window bloat from unused tool prompts - Need to manually disable tools in every agent configuration - No way to set a "baseline" of tools that agents inherit ## Proposed Solution Add a global `defaultTools` array that defines which tools are available to all agents by default: ```json { "$schema": "https://opencode.ai/config.json", "defaultTools": ["read", "grep", "glob", "list"], "agent": { "myagent": { "description": "Custom agent", "tools": { "webfetch": true } } } } ``` ## Behavior • If defaultTools not set: Current behavior (all tools enabled by default) - non-breaking • If defaultTools is set: Only these tools are available by default, agents can add more via their tools config • Users can set "defaultTools": [] (empty array): Zero tools by default, opt-in per agent This allows clean agent contexts without repetitive configuration across all agents.
yindo closed this issue 2026-02-16 17:33:06 -05:00
yindo commented 2026-02-16 17:33:07 -05:00
Author
Owner
Copy Link
Copy Source

@rekram1-node commented on GitHub (Oct 1, 2025):

@calvinalkan there is a tools object at both the agent level and the root of your config.

This means you can disable x or more tools for all agents by default and then enable per agent per tool.

Or you can enable for all agents and then disable per agent

@rekram1-node commented on GitHub (Oct 1, 2025): @calvinalkan there is a tools object at both the agent level and the root of your config. This means you can disable x or more tools for all agents by default and then enable per agent per tool. Or you can enable for all agents and then disable per agent
yindo commented 2026-02-16 17:33:07 -05:00
Author
Owner
Copy Link
Copy Source

@rekram1-node commented on GitHub (Oct 1, 2025):

I think that should allow you to accomplish what you are looking for here right?

@rekram1-node commented on GitHub (Oct 1, 2025): I think that should allow you to accomplish what you are looking for here right?
yindo commented 2026-02-16 17:33:07 -05:00
Author
Owner
Copy Link
Copy Source

@gimbo commented on GitHub (Oct 1, 2025):

@rekram1-node So e.g. this should work to disable all tools by default, including ones provided by MCP servers or new features added over time...?

  "tools": {
    "*": false,
  },
@gimbo commented on GitHub (Oct 1, 2025): @rekram1-node So e.g. this should work to disable all tools by default, including ones provided by MCP servers or new features added over time...? ``` "tools": { "*": false, }, ```
yindo commented 2026-02-16 17:33:07 -05:00
Author
Owner
Copy Link
Copy Source

@calvinalkan commented on GitHub (Oct 1, 2025):

I think that should allow you to accomplish what you are looking for here right?

Oh, i'll check out the source to see how it works. I think it should be what I need.

I think this is missing in the config docs? Or am I blind?

@calvinalkan commented on GitHub (Oct 1, 2025): > I think that should allow you to accomplish what you are looking for here right? Oh, i'll check out the source to see how it works. I think it should be what I need. I think this is missing in the config docs? Or am I blind?
yindo commented 2026-02-16 17:33:07 -05:00
Author
Owner
Copy Link
Copy Source

@calvinalkan commented on GitHub (Oct 1, 2025):

This works for me:

{
  "$schema": "https://opencode.ai/config.json",
  "model": "claude-sonnet-4-20250514",
  "tools": {
    "bash": false,
    "edit": false,
    "glob": false,
    "grep": false,
    "linear_*": false,
    "list": false,
    "lsp-diagnostics": false,
    "lsp-hover": false,
    "multiedit": false,
    "patch": false,
    "read": false,
    "task": false,
    "todoread": false,
    "todowrite": false,
    "webfetch": false,
    "write": false
  },
  "mcp": {
    "linear": {
      "type": "local",
      "command": ["pnpm", "exec", "mcp-remote", "https://mcp.linear.app/sse"],
      "enabled": true
    }
  }
}

Sadly, just doing

tools: {
  "*": false
}

does not work. Looks like the wildcard is not resolved in agent.ts

@calvinalkan commented on GitHub (Oct 1, 2025): This works for me: ```json { "$schema": "https://opencode.ai/config.json", "model": "claude-sonnet-4-20250514", "tools": { "bash": false, "edit": false, "glob": false, "grep": false, "linear_*": false, "list": false, "lsp-diagnostics": false, "lsp-hover": false, "multiedit": false, "patch": false, "read": false, "task": false, "todoread": false, "todowrite": false, "webfetch": false, "write": false }, "mcp": { "linear": { "type": "local", "command": ["pnpm", "exec", "mcp-remote", "https://mcp.linear.app/sse"], "enabled": true } } } ``` Sadly, just doing ```json tools: { "*": false } ``` does not work. Looks like the wildcard is not resolved in agent.ts
yindo commented 2026-02-16 17:33:07 -05:00
Author
Owner
Copy Link
Copy Source

@rekram1-node commented on GitHub (Oct 1, 2025):

@calvinalkan I think we have these settings in a couple places in the docs, thx for pointing out issue with the generic wildcard false that should be working so I guess there is a bug there

@rekram1-node commented on GitHub (Oct 1, 2025): @calvinalkan I think we have these settings in a couple places in the docs, thx for pointing out issue with the generic wildcard false that should be working so I guess there is a bug there
yindo commented 2026-02-16 17:33:07 -05:00
Author
Owner
Copy Link
Copy Source

@rekram1-node commented on GitHub (Oct 1, 2025):

aside from that wildcard not working does this address your issue properly?

@rekram1-node commented on GitHub (Oct 1, 2025): aside from that wildcard not working does this address your issue properly?
yindo commented 2026-02-16 17:33:07 -05:00
Author
Owner
Copy Link
Copy Source

@gimbo commented on GitHub (Oct 1, 2025):

Sadly, just doing

tools: {
"*": false
}
does not work. Looks like the wildcard is not resolved in agent.ts

That's odd, because it does seem to work for me.

E.g. if I have that, and then add another primary agent with no tools specified:

    "trevor": {
      "mode": "primary"
    },

... then it doesn't seem to be able to use webfetch or bash when I try.

This is opencode v0.13.7, hitting amazon-bedrock/anthropic.claude-3-7-sonnet-20250219-v1:0.

@gimbo commented on GitHub (Oct 1, 2025): > Sadly, just doing > > tools: { > "*": false > } > does not work. Looks like the wildcard is not resolved in `agent.ts` That's odd, because it does seem to work for me. E.g. if I have that, and then add another primary agent with no tools specified: ``` "trevor": { "mode": "primary" }, ``` ... then it doesn't seem to be able to use `webfetch` or `bash` when I try. --- This is opencode v0.13.7, hitting `amazon-bedrock/anthropic.claude-3-7-sonnet-20250219-v1:0`.
yindo commented 2026-02-16 17:33:07 -05:00
Author
Owner
Copy Link
Copy Source

@gimbo commented on GitHub (Oct 1, 2025):

In general I think the question of what tools are actually available/unavailable in a given context is a bit opaque. Perhaps if #2666 gets addressed that would help as one could then at least see what information the models are being fed.

Even from what is in the docs, it's hard to discern the rules. E.g. in docs/agents#tools, we have:

If no tools are specified, all tools are enabled by default.

OK, but what if some tools are specified? Are all other tools disabled in that case? That seems like the obvious inference, and I guess it would be sensible — but it's not explicit anywhere, and I haven't been able to establish the answer to this just by trying things out (maybe a skill issue on my part — again, #2666 would help here).

@gimbo commented on GitHub (Oct 1, 2025): In general I think the question of what tools are actually available/unavailable in a given context is a bit opaque. Perhaps if #2666 gets addressed that would help as one could then at least see what information the models are being fed. Even from what _is_ in the docs, it's hard to discern the rules. E.g. in [`docs/agents#tools`](https://opencode.ai/docs/agents/#tools), we have: > If no tools are specified, all tools are enabled by default. OK, but what if some tools _are_ specified? Are all other tools disabled in that case? That seems like the obvious inference, and I guess it would be sensible — but it's not explicit anywhere, and I haven't been able to establish the answer to this just by trying things out (maybe a skill issue on my part — again, #2666 would help here).
yindo commented 2026-02-16 17:33:07 -05:00
Author
Owner
Copy Link
Copy Source

@calvinalkan commented on GitHub (Oct 1, 2025):

Okay, I tested this more deeply, and I can confirm it behaves as follows:

"*": false disables all tools globally, but the built-in plan and build agents still
have some tools enabled by default in the source code (they spread defaultTools but
start with an empty object, not hardcoded tools).

My config:

{
  "tools": { "*": false },
  "agent": {
    "plan": { "tools": { "linear_*": true } }
  }
}

This disables everything NOT explicitly enabled in any agent, then gives plan agent access to
Linear MCP.

So the wildcard works.

@calvinalkan commented on GitHub (Oct 1, 2025): Okay, I tested this more deeply, and I can confirm it behaves as follows: `"*": false` disables all tools globally, **but** the built-in plan and build agents still have some tools enabled by default in the source code (they spread defaultTools but start with an empty object, not hardcoded tools). My config: ```json { "tools": { "*": false }, "agent": { "plan": { "tools": { "linear_*": true } } } } ``` This disables everything NOT explicitly enabled in any agent, then gives plan agent access to Linear MCP. So the wildcard works.
yindo commented 2026-02-16 17:33:07 -05:00
Author
Owner
Copy Link
Copy Source

@gimbo commented on GitHub (Oct 1, 2025):

Another possible way to provide visibility of this would be to explicitly unpack the information in the API server, so that e.g. on the /agent endpoint, instead of (say):

{
  "name": "build",
  "tools": {
    "*": false,
    "write": true,
    "edit": true,
    "bash": true
  },
  ...

(which is what's there now, just combining the config from the top level with the per-agent config)

... we could have (e.g.):

{
  "name": "build",
  "tools": {
    "bash": true,
    "edit": true,
    "glob": false,
    "grep": false,
    "linear_*": false,
    "list": false,
    "lsp-diagnostics": false,
    "lsp-hover": false,
    "multiedit": false,
    "patch": false,
    "read": false,
    "task": false,
    "todoread": false,
    "todowrite": false,
    "webfetch": false,
    "write": true
    },
    ...

... if that makes sense?

@gimbo commented on GitHub (Oct 1, 2025): Another possible way to provide visibility of this would be to explicitly unpack the information in the API server, so that e.g. on the `/agent` endpoint, instead of (say): ``` { "name": "build", "tools": { "*": false, "write": true, "edit": true, "bash": true }, ... ``` (which is what's there now, just combining the config from the top level with the per-agent config) ... we could have (e.g.): ``` { "name": "build", "tools": { "bash": true, "edit": true, "glob": false, "grep": false, "linear_*": false, "list": false, "lsp-diagnostics": false, "lsp-hover": false, "multiedit": false, "patch": false, "read": false, "task": false, "todoread": false, "todowrite": false, "webfetch": false, "write": true }, ... ``` ... if that makes sense?
yindo commented 2026-02-16 17:33:07 -05:00
Author
Owner
Copy Link
Copy Source

@rekram1-node commented on GitHub (Oct 1, 2025):

right now we have a notion of permissions and one for tools, as part of the opentui migration these concepts will be merged and make these spreads and other behaviors both more intuitive and easier to work with (just fyi)

@rekram1-node commented on GitHub (Oct 1, 2025): right now we have a notion of permissions and one for tools, as part of the opentui migration these concepts will be merged and make these spreads and other behaviors both more intuitive and easier to work with (just fyi)
yindo commented 2026-02-16 17:33:07 -05:00
Author
Owner
Copy Link
Copy Source

@calvinalkan commented on GitHub (Oct 1, 2025):

right now we have a notion of permissions and one for tools, as part of the opentui migration these concepts will be merged and make these spreads and other behaviors both more intuitive and easier to work with (just fyi)

IMO, current way is good and flexible, it just needs to be in the docs what the expected bevahiour is.

@calvinalkan commented on GitHub (Oct 1, 2025): > right now we have a notion of permissions and one for tools, as part of the opentui migration these concepts will be merged and make these spreads and other behaviors both more intuitive and easier to work with (just fyi) IMO, current way is good and flexible, it just needs to be in the docs what the expected bevahiour is.
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
acp
bad
bug
discussion
docs
documentation
enhancement
extension
good first issue
growth
help-wanted
mcp
model-problem
needs:compliance
nix
opentui
perf
pull-request
Mirrored from GitHub Pull Request
 tui
web
windows
zen
No Label
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
yindo
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: anomalyco/opencode#1895
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?